Model versions and quantization

Open the download page for almost any local model and you will find not one file but several -- names ending in fp16, fp8, or GGUF Q8, Q4, and so on. They are all the same model. What differs is precision: how much detail is kept in each of the model's numbers. Lower precision means a smaller file that needs less graphics memory and often runs faster, at some cost to quality. This is called quantization.

The practical upshot: a model that "needs a 24 GB card" at full precision will often run happily on a 12 GB card in a smaller version. Picking the version that fits your card is the key skill for running models locally.

One warning about the word "version", which this page uses in a narrow sense. Here it always means which file of one model -- fp16 against fp8 against GGUF. It does not mean which release of a model family, as in Wan 2.1 against Wan 2.2. Those are a different question, covered in Model families and version names.

The versions, from biggest to smallest

• fp16 / bf16 (full precision) -- the original quality, the largest file, and the most graphics memory. Use it if your card has room.
• fp8 (half precision) -- roughly half the memory of fp16 for a small, usually hard-to-notice drop in quality. A good default when the full version is too big. It works with add-on styles (LoRAs) normally.
• GGUF (Q8 down to Q4 and lower) -- a family of compressed versions that let big models run on modest cards. Q8 is close to full quality; the quality cost grows as the number drops, and heavy compression (Q4 and below) can visibly soften results. GGUF files need a GGUF-aware setup to load them.

Which version fits my card?

These are rough starting points, not guarantees -- actual fit depends on the specific model, the resolution you generate at, and what else is loaded:

Graphics memory	Rough guidance
8 GB	Smaller models, or GGUF (Q4-Q6) of larger ones. Video will be a stretch -- consider the cloud.
12 GB	fp8 of most image models; GGUF for the largest ones. Short video is possible for smaller models.
16 GB	fp8 comfortably; fp16 for many image models.
24 GB and up	fp16 of most things, including video.

If your card is below the model you want, you have two easy options: use a smaller version, or run that model on Comfy Cloud instead.

Add-on styles (LoRAs) and quantization

Style add-ons called LoRAs generally work fine on fp16 and fp8. On GGUF versions they still work, but they are applied a little differently under the hood, which costs a bit of speed and can reduce fidelity on the heavily compressed versions. If a LoRA looks weak on a heavy GGUF version, trying fp8 is often the fix.

You usually do not choose by hand

When you run a model locally, the platform aims to pick a version that fits your graphics card, so you rarely select fp8 or Q4 yourself. When you run on Comfy Cloud, it always uses the full-precision version, since the cloud hardware has the memory for it. The main time you think about versions directly is when you already have a particular file on disk and want to use it -- which the Files needed list, in the ComfyUI section of Settings, lets you do.

Quantization methods and the quality they deliver keep improving. Treat the specifics here as a current snapshot and check a model's own notes for anything it recommends.